Continual learning, adaptive skills, and safe runtime improvement

Inference-time adaptation is advancing through GUI skill reflection, recursive research, context compression, memory, and experience-based refinement without necessarily changing base weights. Dynamic behavior interventions create lightweight customization opportunities, but poisoning, forgetting, unsafe updates, alignment regressions, calibration drift, rollback, and reproducibility remain major barriers.